Generalizing from Simple Instances: An Uncomplicated Lesson from Kids Learning Object Categories Ji Y. Son (
[email protected]) Linda B. Smith (
[email protected]) Robert L. Goldstone (
[email protected])
Department of Psychological and Brain Sciences, 1101 E. 10 th Street Bloomington, IN 47401 Abstract Abstraction is the process of stripping away irrelevant information so that learners can generalize on relevant similarities. Can we shortcut this process by directly teaching abstractions in the form of simplified instances? We tested this prediction in the domain of shape-based generalization and found that young children were able to generalize better when taught with simplified shapes rather than complex detailed ones. Simplicity during training allowed shape novices to generalize like shape experts.
Introduction Applying past learning to new circumstances requires the recognition of similarities between those past experiences and the present. The relevant similarities are often embedded with many task-irrelevant similarities and differences. Thus, processes of abstraction – of finding the right similarities – are crucial to theories of generalization (see Harnad, 2005 for a defense of this assumption). Abstraction and generalization are also crucial to understanding the differences between immature and mature learners and between novices and experts, as mature learners generally, and experts more specifically (e.g., Chi, Bassok, Lewis, Reimann & Glaser, 1989; Gick & Holyoak, 1983), seem able to abstract the right similarities over which to generalize past experiences. Such abstracted understandings may be responsible for experts’ ability to transfer their learning to highly dissimilar situations (Holyoak, 1984). The broad goal of the experiments reported in this paper is a deeper understanding of this relationship between abstraction and generalization. If the key to generalization is the formation of a sufficiently minimal description of the relevant properties, then one should be able to directly teach that abstraction and, as a consequence, get broad and appropriate transfer. Some studies with adults learning difficult domains such as chicken sexing (Biederman & Shiffrar, 1987) or scientific principles (Goldstone & Sakamoto, 2003; Sloutsky, Kaminski & Heckler, 2005) have shown generalization benefits when information is presented with more perceptually abstract forms that leave out irrelevant details. In this paper, we ask whether training with abstractions increases transfer in a specific domain: the development of 3-dimensional object recognition in toddlers. Around 2 years of age, when young children become experts in generalizing names for things to new
instances, is this an expertise based on abstraction?
Development of Shape-based Noun Categories Early word learning is defined by proper generalization. Children comprehend that certain nouns go with particular categories of objects at about 9 months of age (Huttenlocher, 1974). Paul Bloom (2000) summarizes how extremely early words are learned and extended to new instances slowly but soon the pace of both learning and generalization accelerates such that shortly after 24 months, children add words to their vocabulary at a staggering rate and also generalize a newly learned name broadly and correctly to category members after experiencing just one instance (Markman, 1989). These young word learners do not need to experience a whole variety of elephants or staplers to know the range of things that are elephants or staplers. One example will do; apparently these children know the right similarity to generalize over. The relevant similarity, at least for concrete noun categories, often involves shape (e.g., Clark, 1973; Imai, Gentner & Uchida, 1994). The key experimental results documenting the importance of shape to children’s noun generalizations derive from a task in which children are taught a name with a single never-before-seen exemplar then asked to generalize that name to new also never-beforeseen instances. In these tasks, 18 month olds show a limited bias to extend object names by shape whereas 30 month olds systematically extend the category name to new instances by shape, ignoring a variety of other properties including color, size, material, and fine-grained details. At the same time children also become able to recognize common object categories from highly minimalist representations of their 3-dimensional shapes. Figure 1 shows an example of a minimalist shape which leaves out finer grained details, coloring, and texture information in contrast to the richly detailed and lifelike versions presented to 18 to 24 month olds in an experiment by Smith (2003). Although to adults these objects seem very similar, younger children with more limited word knowledge did not recognize the simplified forms but did recognize (nearly perfectly) the richly detailed versions when asked, “Where is the ice cream?” In contrast, slightly older children with more advanced word knowledge recognized the simplified shapes just as well as they recognized the richly detailed shapes. Smith proposes that the process of category learning was abstracting shape descriptions.
objects were constructed from 2 to 4 geometric components. They had no details and were painted a uniform color. There were 12 unique objects painted in 12 unique colors, that were arranged into 3 stimulus sets of four objects each. Each set contained (1) a target exemplar trained with a name, (2) an unnamed training distractor, (3) a transfer target with the same exact shape as the exemplar but different in color, and (4) a transfer distractor with the same exact shape as the training distractor but different in color. Figure 2 shows a training exemplar and transfer target pair for the complex condition and for the simple condition. Unique names were paired with each of the three training exemplars: zupp, wazzle, and peema. Figure 1: Realistic detailed stimuli and its corresponding minimalist shape of a familiar object used in Smith (2003). Are simple representations responsible for shape-based generalization? Then these very young children, who do not recognize simple shapes nor generalize by shape, may benefit from explicit training with simple instances rather than complex detailed ones. In particular, simple training should result in shape-based generalization. Our experiments use two types of training: novel names linked to simple objects or complex ones. But there is one driving question: Is there greater generalization from training with the simplified or detailed shape? The results of the two experiments provide insights into the development of object recognition and also into broader issues of learning.
Experiment 1 Children are presented with a single novel exemplar and taught its name and then asked to generalize the name to other objects. Children participated in one of two conditions. In the complex condition, names were paired with richly detailed exemplars. The generalization targets were the very same detailed shape as the exemplar – only differing in color. In the simple condition, the names were paired with simple shape abstractions and the generalization targets were the same shapes in a different color. These simple stimuli provide shape information unencumbered by frills and details. In contrast, the stimuli in the complex condition present lots of extraneous (and potentially useable) features. If minimalist representations of shape promote generalization, and rich ones hamper it, then these young children should show more appropriate patterns of generalization in the simple than in the complex condition.
Procedure and Design. There were two between-subject conditions, Complex and Simple. Children in the complex condition only saw complex objects and those in the simple only saw simple objects. The task was based on one used previously by Woodward and Hoyne (1999). In the familiarization phase, the child was taught the name of the target (e.g., “This is a zupp.”) and also acquainted with a second object, the distractor, that was not named (e.g., “Look at this.”). Objects were present ed one at a time. This familiarization sequence was repeated twice. The second phase, test, occurred after a 3 sec delay. This phase began with a memory test. The original target and distracter were placed on the table and the child was asked to get the target by name (e.g., “Where is the zupp?”). The memory test was immediately followed by a generalization test, two new objects, the transfer target and transfer distractor, were placed on the table, one matching the training target in exact shape, the other matching the training distractor. Both of these generalization objects differed in color from the familiarization exemplar and distracter. The child was asked for the target by name. The memory and generalization tests were then repeated for this same set. The spatial location of the correct choice alternated between test trials. This whole procedure was repeated for each of the 3 unique stimulus sets, yielding a total of 6 memory tests (2 per unique stimulus set) and 6 generalization tests (2 per unique stimulus set).
Method Participants. Thirty-one children participated (15 male, 16 female). The mean age was 17 months (range 15 to 20 months). Materials. Two corresponding sets of novel objects, complex and simple, were created for this task. Fantasy vehicle toys were purchased for the complex objects. They had detailed parts that were intricately painted using three different colors to enhance the finer details. The simple
Figure 2: Exemplar/transfer target pairs for complex (top) and simple (bottom) conditions in Experiment 1.
Results and Discussion Performance on both memorization and generalization trials are shown in Figure 3. Children in both the simple and complex conditions exhibited similar performance on the memory trials, scoring 67% correct, t(29) = -.231. Clearly they can link a name to a specific object, simple or complex, and remember it. In contrast, performance on the generalization tests showed a strong effect of stimulus type. Children correctly generalized the name of the simple exemplars more than complex ones, t(29) = -2.495, p
